Guided Past The Conditional

A conditional diffusion model is sampled with classifier-free guidance, $\tilde\epsilon_\theta(z_\lambda, c) = (1+w),\epsilon_\theta(z_\lambda, c) - w,\epsilon_\theta(z_\lambda)$ .

At one sampling step and one coordinate the conditional prediction is $0.40$ , the unconditional prediction is $0.20$ , and $w = 3$ . What noise value does this step use?

$0.35$ , the $3{:}1$ weighted average of the conditional and unconditional predictions

$1.00$ , from $4(0.40) - 3(0.20)$

$0.80$ , from $w,\epsilon_c + (1-w),\epsilon_u$

$-0.40$ , from $(1+w),\epsilon_u - w,\epsilon_c$